Winter School on Deep Learning
Organised by
Electronics and Communication Sciences Unit
Indian Statistical Institute
Course-End Assessment (Sol.)
Date: March 23, 2024 Duration; 1
1
/
2
hours Points: 42
Name:
Email ID: Registration ID:
Instruction: The test is comprised of 42 questions.
1. All the participants are requested to attempt questions as many as they
can within the stipulated time.
2. No extra time will be provided.
3. For multiple-correct-choice question, you need to answer exactly all the
correct answers. You will be awarded zero otherwise.
4. Participants are prohibited from using any generative agent.
5. Participants must join the zoom meeting link during the assessment
and keep their video on.
6. The Assessment is usually circulated using Google Forms.
Page Points Score
1 4
2 8
3 8
4 7
5 5
6 6
7 4
Total: 42
1. What does the ‘batch size’ parameter in a DataLoader represent?
A. The number of epochs to train the model.
B. The number of samples in each mini-batch.
C. The number of classes in the dataset.
D. The learning rate for the optimizer.
2. What is the fundamental equation governing the residual block of ResNet?
A. y = f(x) B. y = x + f(x) C. y = x ∗ f(x) D. y = x − f(x)
3. What are the two main operations involved in depthwise separable convolution?
A. Strided convolution and dilated convolution.
B. Depthwise convolution and pointwise convolution.
C. Transposed convolution and group convolution.
D. Max-pooling and average-pooling.
4. Which of the following are reasons to use image augmentation techniques?
Increase the size of the validation dataset.
✓ Improve model generalization.
Enhance the resolution of the images.
✓ Reduce overfitting.
5. If an input image of size 32 × 32 is convolved with a 3 × 3 filter using a stride of 2 and no padding,
what is the resulting output size?
A. 15 × 15 B. 16 × 16 C. 17 × 17 D. 18 × 18
6. Which of the following image transformation is least likely to be an adversarial attack on an image
classifiers?
A. JPEG compression.
B. Printing the image on a sheet of paper and scanning it again to an image of the same format.
C. Mixup with another randomly selected image where the mixup coefficient is sampled from a
Beta distribution with parameters (0.1, 0.1).
D. Scaling the image between [−1, 1] and then transferring it back to [0, 255].
7. Which of the following statement(s) are true for large language models?
Autoregressive models cannot consider bidirectional context.
For autoregressive models it is not straightforward to find a feature embedding for an input
sentence.
The next sentence prediction task in BERT did not find uses in later models like RoBERTa.
✓ All of the above.
8. Which of the following is not a text embedding that is fit for downstream learning tasks?
A. GLoVE B. Word2Vec C. Positional Embedding D. TF-IDF
9. Which of the following statement is true for tokenization?
A. BERT uses a BytePairEncoding tokenizer.
B. GPT uses a WordPiece tokenizer.
C. WordPiece only merges two token if their joined frequency is less than that of the component
tokens.
D. If there is no EOD token and PAD is instead used then the LLM may never stop
finishing generation during inference.
10. Which of the following statement(s) are true for adversarial attacks?
✓ Gradient ascent on samples may perform a white box adversarial attack.
✓ One objective of adversarial attack is to preserve perceptual similarity.
✓ Black box adversarial attacks may focus on minimizing the query cost.
None of the Above.
11. Which of the following is true for a k-layered Graph Convolutional Network (GCN)?
A. Aggregates information from the k-hop neighbors.
B. Aggregates information from exactly k neighbors which are k-hop away.
C. Aggregates information from the k-length paths.
D. None of the above.
12. Suppose you are told to classify a set of molecules, then which of the following neural networks will be
most suitable
A. Residual Neural Networks.
B. Graph Neural Networks.
Points earned: out of a possible 8 points
C. Recurrent Neural Networks.
D. Hyperbolic Convolutional Neural Networks.
13. Suppose x
i
∈ {0, 1} for all i ∈ [n]; n ∈ N. Which of the following best suit(s) W and b to represent the
function ∧x
i
using a one-layer neural network:
Recall that, 1
n
is a vector of length n having all values set to 1 and a one-layer neural network is σ
W
⊤
x + b
where W
⊤
is a (1 × n) and x is a (n × 1) vector and b is a scaler. Also, σ(p) = 0 for p < 0; σ(p) = p for 0 ≤ p <
1; σ(p) = 1 for p ≥ 1. Moreover, ∧p
i
signifies the logical and operation over all p
i
’s.
A. (1
n
, −(n − 1)) B. (0.5
n
, 2n) C. (0.5
n
, (−0.5n + 1)) D. (−0.3
n
, −(n − 0.3))
14. Which of the following statements correctly describe the basic formula for calculating the F1 score?
A) F1 score is calculated as the harmonic mean of precision and recall.
B) F1 score is calculated as the arithmetic mean of precision and recall.
C) The formula for F1 score is F1 = 2 × (precision × recall)/(precision + recall).
D) F1 score is calculated as the product of precision and recall.
E) F1 score takes into account both false positives and false negatives.
Please select all correct options:
✓ A ✓ C D ✓ E B
15. Consider a Multi-Layer Perceptron (MLP) model with one hidden layer and one output layer. The
hidden layer has 10 neurons, and the output layer has 4 neurons. The input to the MLP is a 4-
dimensional vector. Each neuron is connected to every neuron in the previous layer, and a bias term
is included for each neuron. The activation function used is the sigmoid function. Calculate the total
number of trainable parameters in this MLP model.
A. 95 B. 93 C. 94 D. 150
16. Which of the following methods are commonly overridden when subclassing nn.Module in PyTorch?
✓ init() ✓ forward() backward() parameters()
17. Which of the following statements about nn.Module in PyTorch is/are true?
✓ nn.Module can be used to define both neural network layers and entire models.
✓ Subclassing nn.Module allows for automatic differentiation.
✓ nn.Module provides methods for managing model parameters.
nn.Module is only used for defining custom loss functions.
18. Which of the following is not a typical usecase for the nn.Module class in PyTorch?
A. Implementing a convolutional neural network (CNN).
B. Defining a custom loss function.
C. Creating a recurrent neural network (RNN).
D. Loading pre-trained models for transfer learning.
19. Which of the following is a common technique used to prevent overfitting in deep neural networks?
✓ Dropout. ✓ Batch normalisation. Change activation function. Test on training data.
20. Suppose we define two tensors t1 and t2 as follows:
t1 = torch.randn(size=(2, 3, 4, 9))
t2 = t1.view(6, -1, 3)
What is the value of t2.size()[1]?
Simply write the value or write ‘Error’ if you think this code is not valid.
20.
12
Points earned: out of a possible 8 points
21. In a deep learning model with 8 hidden layers, each containing 200 neurons, a dropout rate of 0.25 is
applied during training. If the model is trained on a dataset consisting of 10, 000 samples, how many
neurons, on average, will be active (not dropped out) during the forward pass of each sample?
21.
1200
22. What should the bonuses be for the company’s employees, Alice and Bob, given the following profit
distribution scenarios? (v is the profit calculation function.)
- When there are no employees, resulting in no profit (v({}) = 0).
- When Alice alone generates 20 units of profit (v({Alice}) = 20).
- When Bob alone generates 10 units of profit (v({Bob}) = 10).
- When both Alice and Bob together generate 50 units of profit (v({Alice, Bob}) = 50).
Hint: Use the concept of Shapely values.
A. Alice Bonus: 20 ; Bob Bonus: 30;
B. Alice Bonus: 40 ; Bob Bonus: 10;
C. Alice Bonus: 30 ; Bob Bonus: 20;
D. None of the above.
23. Consider the 3 × 3 matrix A whose eigenvalues are 2, −2 and −1. Then Trace of (A
3
+ A
2
) is:
A. Not Defined. B. 8. C. 16. D. Invalid Question.
24. Euclidean distance based k-means clustering algorithm was run on a dataset of 100 points with k = 3.
If the points (1, 1) and (−1, 1) are both part of cluster 3, then which one of the following points is
necessarily also part of cluster 3?
A. (0, 0) B. (0, 2) C. (2, 0) D. (0, 1)
25. During WGAN training on a dataset of handwritten digits, you observe:
- Discriminator loss remains high and plateaus.
- Generator loss continues to decrease.
Assuming a fixed model architecture:
✓ The discriminator is too powerful. Reduce its learning rate.
The generator is learning artifacts. Increase its learning rate.
The models are stuck in a local minimum. Adjust the weight clipping range.
More data is needed to improve both losses.
26. Consider a Wasserstein GAN (WGAN) with a gradient penalty term (λ×∥∇D(x)∥
2
) in the discriminator
loss. Which of the following are true:
✓ The gradient penalty encourages the discriminator map to follow Lipschitz conti-
nuity with a constant of 1.
The gradient penalty encourages the discriminator to minimize the Jensen-Shannon divergence
between real and generated data distributions.
✓ A very large λ can lead to vanishing gradients in the discriminator, hindering
training progress.
An optimal λ value ensures the discriminator perfectly distinguishes real from generated sam-
ples.
27. What information does the score function (∇S(x
t
)) provide about the data point (x
t
)?
Points earned: out of a possible 7 points
✓ The score function indicates the direction of steepest ascent in the data distribu-
tion’s probability density at point x
t
.
A higher score signifies that the data point is closer to a clean data point (x
0
) with less noise.
The score function directly estimates the amount of noise present in the data point (x
t
).
✓ By following the negative gradient of the score function, the model can iteratively
remove noise and approach the clean data distribution.
28. Which of the following represents the Betti numbers (β
0
, β
1
, β
2
) of the given figure (A Hollow Torus)?
A. (2, 1, 0). B. (1, 2, 1). C. (1, 1, 0). D. (0, 2, 1).
29. Which statement correctly distinguishes between the Vietoris–Rips and
ˇ
Cech filtrations in Topological
Data Analysis?
A. The Vietoris–Rips complex depends on the intrinsic geometry of the dataset, while the
ˇ
Cech
complex relies on an embedding into a larger space.
B. The Vietoris–Rips complex considers pairwise distances, whereas the
ˇ
Cech com-
plex involves all possible combinations of intersections of balls.
C. The Vietoris–Rips filtration is used for communication coverage analysis, while the
ˇ
Cech fil-
tration provides persistence information.
D. The Vietoris–Rips complex is a subcomplex of the
ˇ
Cech complex.
30. The maximum number of full-dimensional Linear Regions resulting from partitioning defined by n
hyperplanes of the dimension d(d < n) is:
✓ 2
n
2
d
n
P
i=1
d
i
d
P
i=1
n
i
31. Let A, B be (2 × 2) matrices entries from Real numbers such that |A| = |B| and Tr(A) = Tr(B). Then,
which among the following is an incorrect option?
A. A and B have the same set of eigenvalues.
B. A and B need not be similar.
C. A and B are similar.
D. None of these.
32. What is the output of the following MP (McCulloch-Pits) neuron for the AND Boolean function?
y =
(
1, if x
1
+ x
2
+ x
3
≥ 1
0, otherwise
✓ y = 1 for (x
1
, x
2
, x
3
) = (0, 1, 1) y = 0 for (x
1
, x
2
, x
3
) = (0, 0, 1)
✓ y = 1 for (x
1
, x
2
, x
3
) = (1, 1, 1) y = 0 for (x
1
, x
2
, x
3
) = (1, 0, 0)
Points earned: out of a possible 5 points
33. How many times WSDL will be printed when you execute this simple python coding?
i=0
while True:
i += 1
if i > 32:
break
elif i<= 10:
print('WSDL')
continue
print('ISI')
print(i)
A. 10. B. 11. C. 31. D. 32.
34. Why do Markov Decision Processes (MDPs) form a base for any Reinforcement Learning (RL) prob-
lem?
Because MDPs ensure that the rewards are deterministic.
✓ Assuming the future state depends only on the current state and action and not
on the sequence of events that preceded it.
Eliminating the need for an action space.
By ensuring that the optimal policy can be found without knowing the state transition prob-
abilities.
35. In the context of RL, which of the following statement(s) best describe the function of the discount
factor (γ)?
✓ It determines the present value of future rewards.
✓ A value of γ close to 1 encourages the agent to consider long-term rewards more
strongly.
It has no impact on policy optimization in deterministic environments.
✓ A value of γ equal to 0 makes the agent myopic, focusing only on immediate
rewards.
36. Which of the following statement(s) is/are true regarding the exploration-exploitation dilemma in RL?
Pure exploration regards the knowledge gained from past experiences.
✓ Exploration is necessary for discovering optimal policies.
✓ Exploitation is solely concerned with maximizing immediate gains.
✓ A balance between exploration and exploitation is often achieved through strate-
gies like ε-greedy.
37. What is meant by a policy (π) in RL?
The set of all possible actions in an environment.
The method for updating the Q-function based on new information.
✓ The behavior of an agent in terms of action selection in each state.
The maximum reward that can be obtained from a given state.
38. For Deep Q-Networks (DQN), which of the following technique(s) contribute to stabilizing the training
process?
✓ Replay Buffer. Gradient Clipping. Regularisation. ✓ Target Networks.
Points earned: out of a possible 6 points
39. Which of the following is the main splitting criterion for the CART algorithm?
A. Gain Index. B. Information Gain. C. Divergence Gain. D. Gini Impurity.
40. Using the concept of broadcasting in PyTorch, predict the output of the following code snippet.
x=torch.ones(5,2,1,7)
y=torch.ones(3,1,1)
print((x+y).size())
torch.size([5,1,4,3]) ✓ Error torch.size([5,3,4,1]) None of the Above
41. Which of the following options is incorrect about the computational graph in PyTorch auto-grad?
A. Leaves of the Computational Graph are the input tensors and roots are the output tensors.
B. The computational graph is a directed graph.
C. There may be cycles in the in a computational graph.
D. None of the Above.
42. Which of the following is not an attribute of the torch.Tensor type of variable?
A. torch.dtype. B. torch.device. C. torch.shape. D. torch.sparse layout.
Points earned: out of a possible 4 points
